Fathom: artificial intelligence at the USB port

Fathom is a USB device with a Myriad 2 processor capable of running computer vision algorithms in real-time. This little device has the potential to change the applying mode of artificial intelligence in robotics.

Let suppose that you want to build a robot capable of recognizing objects and sort them according to the features, or search a certain toy somewhere in the house, or all of these together. I speak from experience here. It is damn hard to do all of these without an algorithm to recognize objects in real time, and of course, a lot of late nights of work.

First of all, for Computer Vision (CV) applications you need a powerful computer. From the hardware side, the Raspberry Pi 3 board is widely used in such applications. From the software point of view, you need a library of video/image processing such as OpenCV, in some cases a cloud account, and certainly a ton of program lines.

Google made this work a little easier and developing an application called TensorFlow. But even so, it is still hard to make a robot recognize or distinguish objects based on computer vision algorithms.

Fathom by Movidius

Fathom by Movidius

Let get back to our intelligent device and discover what is able to do.

Fathom is a USB gadget that fits into any device with USB ports. A list of such devices is given even to the device’s manufacturer. The list includes hobby platforms such as Raspberry Pi and Arduino, the GoPro cameras, and of course any computer. Immediately after installation, any of the above devices are capable of using the power of the Myriad 2 processor and neural networks.

A version of a neural network already comes installed with the Fathom device. In other words, if you want to make a robot to make decisions based on images captured by a video camera, you do not need OpenCV or computer vision algorithms in the cloud.

All that means “Computer Vision” applications eat a lot of hardware resources and energy. If you use CV for a drone or a mobile robot, the power consumption is very important. In this case, we have to take into account the power consumption for every component. What is to note here, is that the Myriad 2 processor has a very low power consumption. About the consumption details, I will talk later in the article.

Writing the code
Movidius, the company that launched Fathom, created a Myriad Development Kit (MDK). The MDK allows the users to develop, compile their own programs, and use software specifically designed to help the development of algorithms. The framework is called Deep Fathom Learning Software Framework and is designed to compile algorithms and translate them into a language understood by the Myriad 2 processor.

Myriad 2

Deep Learning on the Myriad platform

Deep Learning on the Myriad platform

Myriad 2 is the brain that makes things happen. The processor runs at 600MHz, which does not seem much for an artificial intelligence system. For a higher processing power, can be chosen the turbo variant that increases processor speed over the normal speed. But this is another topic that I will not discuss here.

Myriad 2 is able to do a number of two trillion operations every second. These operations are true only if we speak of 16-bit processing. The power consumption is 500 milliwatts (0.5 watts). I used temperature sensors with almost the same energy consumption. The power consumption is very low considering that the processor can simultaneously analyze inputs from multiple cameras.

Strictly speaking of analysis in computer vision, Myriad 2 can process 16 images per second.

The processor architecture is specially designed for computer vision and neural network applications. This is not common for all the processors. For example, the Raspberry Pi’s GPU can cover applications of computer vision, but resources consumed are much higher compared to a dedicated processor for this kind of applications.

Other details
The Fathom device is not widely available yet, but the company hopes to launch the product on the market by the end of 2017. Regarding the price, the target is a price of less than $100.

And a video presentation of the project:

Leave a Reply

Required fields are marked *.